Perplexity AI taught its model to self-correct errors
Perplexity AI introduced research findings on improving the training of its Large Language Models (LLMs) by teaching them to correct their own errors independently.
For this, the method of prompt-guided self-distillation and selective learning was used. The essence of the method is that the model learns from its unsuccessful sessions, not only from successful ones, which allows it to identify and correct failures that go beyond standard scenarios.
As a result of testing, the later version of the model showed a significant increase in reliability: the number of tool-calling failures decreased by 21.2%. Furthermore, in a separate test, the model using prompts for error correction avoided the initial error in 93.7% of cases, which is significantly higher than 75.1% for the model without prompt assistance.
Why it matters
- —Increased reliability: The model started making significantly fewer mistakes, especially when using external tools.
- —New learning approach: The methodology of training on its own errors (self-distillation) increases the quality and stability of AI.
- —Practical application: Reducing failures when calling tools is critically important for the real-world use of AI in complex tasks.
Key facts
- Reduced tool-calling failures by 21.2% in an A/B test.
- The model using prompts avoided errors in 93.7% of cases.
- In a live test, tool-calling failures decreased from 2.24% to 1.77%.
- Selective learning methods and prompt-guided self-distillation were used.
The full text is in the original source. Here we provide a brief summary and key facts.